Music Similarity Evaluation Based on Onsets
This paper describes a music similarity approach based on the time differences between two adjacent onsets. To better detect onsets, temporal and spectral detection methods are employed. Each set of detected features are individually matched by using the
- PDF / 298,664 Bytes
- 11 Pages / 439.37 x 666.142 pts Page_size
- 62 Downloads / 203 Views
Abstract This paper describes a music similarity approach based on the time differences between two adjacent onsets. To better detect onsets, temporal and spectral detection methods are employed. Each set of detected features are individually matched by using the rough longest common subsequence (RLCS) algorithm. The final score is a weighted sum of individual scores from each detection method. The simulation results show that, on the average, 85% of the audiences agree that two musical soundtracks are similar if the computed score is greater than 0.3. When compared with an existing approach, it is easier for the proposed approach to set up a threshold to recommend highly similar soundtracks. Keywords Music similarity
⋅
Onsets
⋅
Rough longest common subsequence
1 Introduction With the advances of technology, more and more online soundtracks are available to music lovers. For many instances, a music lover may want to listen to more soundtracks similar to his/her favorite ones. To provide this type of service, techniques for music recommendation can be applied. Currently, there are two approaches to provide the recommendation list. The first one is based on the preference of other users, whereas the second one is based on the temporal and/or spectral similarity of the soundtracks. The first approach is easy to implement. For example, if two soundtracks A and B are frequently downloaded or listened by many users together, then we may assume that A and B are similar. Therefore, if the user requests to recommend soundtracks similar to A, then soundtrack B will be recommended. Though effective, this approach, nevertheless, does not truly recommend “similar” soundtracks to the query soundtrack. Furthermore, this approach S.D. You (✉) ⋅ R.-W. Chao Department of Computer Science and Information Engineering, National Taipei University of Technology, Taipei, Taiwan e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2018 N.Y. Yen and J.C. Hung (eds.), Frontier Computing, Lecture Notes in Electrical Engineering 422, DOI 10.1007/978-981-10-3187-8_16
153
154
S.D. You and R.-W. Chao
almost always recommends most popular soundtracks at the time of query, and ignores any really similar ones with only a few downloads (or browse). Finally, this kind of approach requires Internet connection, which is inconvenient for some situations. As the second approach assesses the temporal and/or spectral similarity between two soundtracks, the similarity is truly based on the contents of the soundtracks without referring to other users’ preferences. For this type of approach, we could either provide a set of musical works to train the similarity evaluation system. Or, we could alternatively use a pre-defended metric to measure the similarity between two soundtracks. In this paper, we only consider approaches without prior training for its ease to use. According to Wikipedia [1], there are many different criteria to assess whether two pieces of music are similar, such as based on pitched similarity, non-pitched similarity, an
Data Loading...
